These labels describe different parts of a system. An assistant may use a knowledge base. An agent may trigger a workflow. A CRM automation may contain no generative AI at all. The practical question is not “Do we need AI?” but “What job, information, authority and checking process does this system need?”
AI assistant
Plain English: software you interact with to draft, summarise, search, classify or answer questions. It normally waits for a person to ask and returns something for that person to use.
SME example: A service manager asks for a draft reply to a customer complaint, then checks the facts and sends it.
Decision question: Is the output advice for a person, or will the software take action itself?
Risk note: A friendly conversational interface can make uncertain output sound authoritative. Treat fluency and factual reliability as separate qualities.
AI agent
Plain English: software given a goal, tools and some freedom to choose its next steps. It may inspect records, decide which tool to use and perform several actions before returning a result.
SME example: An enquiry agent reads a web form, checks service area and calendar availability, updates the CRM and prepares a booking—within defined limits.
Decision question: What may it read, decide and change without approval?
Risk note: More autonomy increases the number of ways a small error can propagate. Start with narrow permissions, transaction limits and a clear stop condition.
Workflow automation
Plain English: a repeatable sequence of triggers, rules and actions. Traditional automation follows specified logic; AI can be added where the input is messy, such as interpreting an email.
SME example: When a signed quote arrives, create a job, notify operations and schedule a deposit reminder. AI is optional if every field is structured.
Decision question: Can ordinary rules solve the problem more predictably?
Risk note: Automating a broken process makes errors faster. Map exceptions and ownership before adding software.
Retrieval-augmented generation (RAG)
Plain English: a process that first retrieves relevant material from approved sources, then gives that material to a generative model to help compose an answer.
SME example: A staff assistant retrieves the current returns policy and the customer’s order details before drafting a response.
Decision question: Can the user see which source passages supported the answer?
Risk note: RAG can improve grounding but does not guarantee truth. The wrong document may be retrieved, an old version may rank highly, or the model may misread the evidence.
Knowledge base
Plain English: the maintained collection of information a team or system is allowed to rely on—policies, product details, procedures, approved answers and source records.
SME example: One controlled library holds the current price list, installation guide, warranty rules and escalation contacts.
Decision question: Who approves, dates, retires and replaces each source?
Risk note: A folder full of documents is not automatically a trustworthy knowledge base. Ownership, version control and access restrictions matter.
CRM automation
Plain English: rules or AI-assisted steps that create, update or act on customer relationship management records.
SME example: A new enquiry is deduplicated, assigned to the right territory, given a follow-up date and flagged if consent information is missing.
Decision question: Which field is the system of record, and what happens when two systems disagree?
Risk note: Poor matching can merge two people, overwrite a good record or trigger an inappropriate message. Test duplicates and edge cases, not only the happy path.
Human-in-the-loop
Plain English: a named person reviews, approves, corrects or can stop an automated step at a meaningful point—not merely after the consequence has occurred.
SME example: AI drafts a supplier-risk summary, but a manager must inspect cited evidence before changing payment terms.
Decision question: Does the reviewer have enough time, context and authority to reject the output?
Risk note: A compulsory click is not effective oversight. Reviewers need visible evidence and manageable volumes.
Evaluation
Plain English: a planned test of whether the system performs its intended job, including quality, safety, reliability and business impact.
SME example: Before launch, test 100 representative enquiries and record correct routing, unsupported claims, privacy failures, escalation rate and time saved.
Decision question: What measurable result would justify launching, pausing or rolling back?
Risk note: A polished demonstration is not an evaluation. Use real-world variations, known difficult cases and repeat tests after model, prompt or source changes. NIST’s voluntary AI RMF treats measurement and ongoing management as core parts of trustworthy AI risk practice.
Hallucination
Plain English: generated content that is unsupported, wrong or invented, even though it may read confidently.
SME example: An assistant invents a product warranty term because similar language appeared in its training patterns.
Decision question: Which claims must be verified against a source before anyone relies on them?
Risk note: The control should match the consequence: source citations for routine answers, human approval for important decisions, and no generative step where a deterministic lookup is safer.
Audit log
Plain English: a protected record of significant events: who or what accessed information, which action occurred, when it happened and whether it succeeded.
SME example: The record shows that an agent changed a lead owner at 10:42, the rule it applied and the staff member who later corrected it.
Decision question: Could you reconstruct an incident without relying on memory?
Risk note: Logs can themselves contain personal or confidential information. Limit access, define retention and avoid recording secrets or unnecessary full-text content.
Data minimisation
Plain English: collect, disclose and retain only the information genuinely needed for the defined job.
SME example: A booking assistant needs service type, suburb and preferred time—not a customer’s full historical account notes.
Decision question: If this field were removed, would the task still work?
Risk note: The OAIC advises organisations to consider whether personal information is necessary before entering it into publicly available generative AI tools, and to avoid sensitive information. Australian privacy obligations depend on the organisation and context; obtain appropriate advice rather than assuming a tool’s settings resolve compliance.
A practical decision path
- Name the job. Write the outcome and the current process in one paragraph.
- Choose the simplest mechanism. Use rules for predictable inputs; an assistant for drafting; RAG for approved sources; an agent only when multi-step autonomy is justified.
- Set authority. List permitted data, tools, actions, limits and escalation points.
- Build the evidence layer. Define source ownership, audit events and a representative evaluation set.
- Match review to consequence. Place human approval before irreversible, financial, legal, safety or customer-impacting actions.
- Minimise data. Remove fields, permissions and retention that the job does not require.
- Launch narrowly. Monitor exceptions, re-test after changes and keep a rollback path.
If you have a defined use case and need practical implementation help for an Australian small business, the linked service overview explains the separate advisory and implementation context.
Cyber-security reminder: The Australian Signals Directorate’s guidance for engaging with AI recommends understanding how providers handle data, applying access controls and considering supply-chain and model-related risks. Treat an AI connection as a system integration, not just a new user interface.